Distributed Model Predictive Control based on Dual Decomposition with Neural-Network-based Warm Start

Abstract

This work deals with the application of neural networks to speed up the convergence of a distributed model predictive control (DMPC) algorithm based on dual decomposition. While dual decomposition methods are known to converge to the centralized MPC solution, numerous iterations may be required before convergence is attained, thus increasing computation and communication burden. In this paper, a database containing system states and optimal Lagrange multipliers is created offline to train a neural network, which is incorporated into the online operation of the distributed system. Numerical results on an input-coupled 16 tanks benchmark are provided.